Francisco Sandoval Hernández

Genetic Algorithms on LAN-message Passing Architectures using PVM: Application to the Routing Problem

In this work we address the Genetic Algorithm parallelization problem, over a LAN-based message passing computer architectures, using PVM 3.1 (Parallel Virtual Machine — Public Domain software) as a software integration tool. The strategy used has been to split the problem into independent functions, most of them running in server processors which perform the work, and report periodically their partial results to a master node which redistributes the information in order to improve the work in each server. The strategy we present here handles all data communications through sockets via PVM calls, in such a way that the communication latency and the overall data-passing load are significantly reduced. In addition, a dynamic load balancing and fault tolerant capabilities are achieved. As an application, we study the Routing problem, which is a classical optimization problem with combinatorial complexity.

A curvature-based multiresolution automatic karyotyping system

This paper presents a method to segment, characterise and pair a set of chromosomes in a cell of an eukaryotic organism. This method yields several new features: (i) chromosomes are captured at non-uniform resolution to minimise the problem instance; (ii) segmentation is adaptively conducted by means of a hierarchical structure in a fast way; (iii) the curvature of each chromosome is studied at high resolution by means of attentive steps; (iv) a very short and uncorrelated feature vector is extracted from curvature by analysing its spectral components; and (v) a multistage benchmark classifier is used to pair chromosomes according to shape and banding. The method has been tested with publicly available databases. Results were successfully compared to manual karyotypes.Abstract. This paper presents a method to segment, characterise and pair a set of chromosomes in a cell of an eukaryotic organism. This method yields several new features: (i) chromosomes are captured at non-uniform resolution to minimise the problem instance; (ii) segmentation is adaptively conducted by means of a hierarchical structure in a fast way; (iii) the curvature of each chromosome is studied at high resolution by means of attentive steps; (iv) a very short and uncorrelated feature vector is extracted from curvature by analysing its spectral components; and (v) a multistage benchmark classifier is used to pair chromosomes according to shape and banding. The method has been tested with publicly available databases. Results were successfully compared to manual karyotypes.

Adaptive Fovea Structures for Space-Variant Sensors

In this paper we describe the architecture and data structure of space-variant sensors with reconfigurable cartesian geometries. The ability of these sensors to change the position and size of their high resolution regions or electronic foveas, makes them suitable to compensate the limited performance or coarse fixation characteristics of the mechanical systems utilized for gaze tasks in active vision applications where size, weight or cost could be conditioning factors to the performance or feasibility of the whole system. An alternative to the implementation of these sensors is based on off-the-shelf CCD cameras and devices with reconfiguration capabilities, such as FPGAs. In this way, besides the multiresolution data output, sensor reconfiguration systems let generate additional data adapted to the functions of the higher level modules of the active vision systems. As a result of this computing capability at the sensor level, it is possible to unload the processing stages of certain tasks without penalty in time or significant addition of hardware. An approach to selective foveation tasks and motion detection is presented.

Short-Term Peak Load Forecasting: Statistical Methods Versus Artificial Neural Networks

Two practical techniques: Time Series (TS) and Artificial Neural Networks (ANN), for the one-step-ahead short-term peak load forecasting have been proposed and discussed in this paper. We use weather variables since it is well known that better forecasting performances can be obtained taking them into account. The order selection of TS and the number input neurons of the ANN have are based on the computation of correlation functions. Their performances are evaluated through a simulation study. An extensive test activity of the two techniques shows that have better forecasting accuracy and robustness ANN models.

Genetic Synthesis of Discrete-Time Recurrent Neural Network

In this paper, we proposed a different genetic model for optimizing both network architecture and connection weights of Discrete-Time Recurrent Neural Networks in evolutionary process. Empirical studies show that our model can efficiently generate the appropiate network size and topology for small applications. We have used two experiments: a parity function and a finite state machine for detection of sequences.

Vision-Based walking parameter estimation for biped locomotion imitation

This paper proposes a new vision-based system that can extract walking parameters from human demonstration. The system uses only a non-calibrated USB webcam connected to a standard PC, and the human is only required to put three color patches on one of his legs and walk roughly in a perpendicular plane with respect to camera orientation. The walking parameters are then extracted in real time, using a local tracking system to follow the markers and a fast decision layer to detect the main features of the leg movement. As only one leg can be tracked properly using only one camera, we assume symmetric movement for left and right legs. Once extracted, the parameters have been successfully tested by generating walking sequences for both simulated and real Robo-Erectus humanoid robots.

Generalization of Shifted Fovea Multiresolution Geometries Applied to Object Detection

This work describes a foveal vision system applied to object detection. The novelty of this system consists of carrying the detections using a generalization of the multiresolution shifted fovea images. The main advantage introduced is the great increase of the number of fovea positions allowed in shifted-fovea systems already implemented: this means that the maximum error of placement is reduced to one pixel, implying that any object could be examined at the highest resolution available regardless of its coordinates. The concept is based on increasing the degrees of freedom and the related number of configuration parameters and the application of a new shifting algorithm which allows a higher number of fixation points on the scene and, therefore, reduces the error of fovea positioning on the region of interest and aproaches closer to the required scene details. Besides, we introduce the multiresolution data structure to manipulate and process this type of foveal geometries, as well as the results obtained after applying hierarchical algorithms for segmentation and detection of objects within this type of multiresolution images.

Use of Genetic Algorithms in Neural Networks Definition

At present, there is not a general methodology for neural network definition. In this work we propose an algorithm highly inspired on biological concepts for generating neural networks oriented to solve particular problems given on terms of input and output. With this algorithm we pretend to specify formal tools of general use for network definition, and to disclose underlying processing structures of the living organisms.

Continuous-State Hopfield Dynamics Based on Implicit Numerical Methods

A novel technique ispres ented that implementscon tinuousstate Hopfield neural networks on a digital computer. Instead of the usual forward Euler rule, the backward method is used. The stability and Lyapunov function of the proposed discrete model are indirectly guaranteed, even for reasonably large step size. This is possible because discretization by implicit numerical methodsinheritsthe stability of the continuoustime model. On the contrary, the forward Euler method requiresa very small step size to guarantee convergence to solutions. The presented technique takes advantage of the extensive research on continuous-time stability, asw ell asrecen t resultsin the field of dynamical analysisof numerical methods. Also, standard numerical methods allow for synchronous activation of neurons, thus leading to performance enhancement. Numerical results are presented that illustrate the validity of this approach when applied to optimization problems.

Neural Networks for Contingency Evaluation and Monitoring in Power Systems

In this paper an analysis of the applicability of different neural paradigms to contingency analysis in power systems is presented. On one hand, unsupervised Self-Organizing Maps by Kohonen have been implemented for visualization and graphic monitoring of contingency severity. On the other hand, supervised feed-forward neural paradigms such as Multilayer Perceptron and Radial Basis Function, are implemented for severity numerical evaluation and contingency ranking. Experiments have been performed with successfully result in the case of Kohonen and Multilayer Perceptron paradigms.